UX CASE STUDY · AI · May 2024 - June 2025

Cutting court transcript turnaround 90% with offline AI under strict government privacy laws.

External contractors were drowning in manual audio work and delays stretched to weeks. I designed and built a fully offline AI transcriber that runs on court hardware, generates same-day dirty transcripts, and opens a direct revenue stream without adding headcount.

SPEC · SAT001
Role
UX Designer & Developer
Timeline
May 2024 - June 2025
Stack
Python · WhisperX · Figma
Constraint
100% OFFLINE (NO CLOUD)
TURNAROUND (DAYS)AFTER vs BEFORE 14→0.5
MY ROLE002

Word about my Transcript Workflow Manager got around the courthouse, and court leadership contacted me directly to build an AI transcription tool. I owned the research, the ML pipeline, and the interface design.

Team
Solo design and build
Stakeholders
Court leadership, clerks, external contractors
FIG. 01May 2024 - June 2025
PROBLEM → SOLUTION003

The Problem vs. The Solution

Courtroom audio was a black box. Over 100 hours of unindexed recordings sat in storage, clerks manually scrubbed through tapes, and requesters paid $9.25 a page for transcripts that took days to arrive.

BEFORE003.A

The transcription backlog

Court clerks waited days, sometimes weeks, for external contractors to manually transcribe audio. Every cloud-based AI tool that could have helped was legally off-limits.

P-01
Third-party delaysStrained external contractors manually transcribed audio after court sessions, causing error-prone delays of several days to weeks.
P-02
Strict cloud restrictionsGovernment security and privacy policy prohibited commercial cloud-based AI tools like standard ChatGPT. Nothing on the market was usable.
P-03
No internal capabilityThe courts offered no first-party transcription service. Costs soared to $9.25 a page and clients waited indefinitely for preliminary data.
AFTER003.B

Offline AI automation

A Python desktop app that runs entirely on local court hardware, fits how clerks already work, and outputs a timestamped transcript in hours. No internet required.

S-01
Secure offline pipelineBuilt a Python app on WhisperX with fully offline language models, guaranteeing 100% compliance with local infrastructure rules.
S-02
Speaker diarizationIntegrated machine learning to automatically differentiate and label multiple speakers in chaotic court recordings.
S-03
Same-day delivery and revenueOutputs a time-accurate dirty transcript within hours, creating a direct monetization path for the government without hiring more staff.
THE IMPACT004

The Impact

Going offline-first was the unlock. Cutting out third-party vendors freed budget, earned trust from leadership, and gave the courts a way to generate revenue on their own terms.

Faster delivery
0%
Hours instead of weeks
Security compliance
0%
Zero cloud data transmission
New revenue stream
+0
Direct client monetization

Transcription turnaround time

Days required to generate a preliminary transcript. Manual vendors vs. the offline AI transcriber.
KEY DECISIONS005

Key Design Decisions

Three calls that shaped the system, each forced by a legal, trust, or business constraint rather than a preference.

DEC-01
Fully offline models only

The whole pipeline runs on local court hardware with offline WhisperX models. No byte ever leaves the device.

Why

Government privacy policy prohibited every commercial cloud AI tool on the market. Running local was the only legal path, and it removed the security review question entirely.

DEC-02
Clerk verification before anything is final

A verification layer lets clerks review AI output before a transcript goes anywhere.

Why

Courtroom audio is messy and AI output cannot be treated as gospel in a legal setting. Clerks owning the final check is what made the tool trustworthy enough to use.

DEC-03
Sell dirty transcripts, not certified ones

The product is a rough, timestamped same-day transcript offered at a much lower price than vendor transcripts.

Why

Vendors were permanently booked and charged $9.25 a page. A cheap preliminary transcript could ship in hours and become court revenue without hiring anyone.

DESIGN EVOLUTION006

From Wireframe to Running Portal

The interface started as three wireframe concepts in Figma, judged against how clerks already search, then became a running desktop prototype.

Three concepts before a line of code.

The desktop interface was laid out in Figma first. Each concept was judged against one bar: a clerk should be able to find a hearing, run a transcription, and read the output without a training manual.

  • Familiar search patterns over novel layouts
  • One screen per job: find, transcribe, read
  • Layouts tested against real clerk workflows
LIVE PROTOTYPE007

Try the AI Transcriber

This is the real thing. Load a file, watch it process, and explore the full output interface, exactly as it was designed and shipped.

aitranscriberprototype.kashb.ca Open in new tab ↗
REFLECTION008
My deep dive into AI, and into selling the work

This project taught me AI development from the inside: tensors, transformers, beams, weights, and tuning parameters for courtroom acoustics. Testing meant live court sessions with clerks reading Casablanca into the record. It also taught me to present, since I spent as much time demoing to stakeholders and contractors as I did tuning models. The Flutter interface never shipped because the project paused while HQ made decisions, and I learned that in government work, the pitch and the pipeline matter equally.

Available for Work

Let's build something impactful.

I'm currently open to new opportunities in UX/UI Design, Product Design, and GovTech transformation roles.